The process of validation and refinement of deforestation alerts includes automatic steps and operational steps involving specialists in remote sensing with knowledge about the deforestation dynamics associated with each Brazilian biome.
The automatic steps are responsible for discarding deforestation alerts with little chance of being true (e.g. polygons detected over forestry areas) and for transferring the data needed for the refinement process between the different platforms used by the project (e.g. Google Cloud Platform, Google Cloud Storage and Google Earth Engine).
The experts are responsible for evaluating the alerts produced by the monitoring systems and thus activating the Planet images for the regions to be analyzed and also for spatially refining each of these real deforestation alerts, using the activated Planet images, with spatial resolution of 3-5 meters. Below is a description of each step of this process.
- Data organization: Alerts generated by different platforms (in the Beta phase: SAD, DETER and GLAD) are downloaded and organized monthly in the database where the different auxiliary databases are also connected (e.g. CAR, Conservation Units, Indigenous Lands and other territorial boundaries).
- Pre-processing: Initial stage responsible for selecting which deforestation alerts can be considered valid, taking as reference the monitoring systems of each biome, as well as their respective classes of native vegetation cover, according to the MapBiomas map for 2017 (see table below). At this stage, some alerts can be ignored (e.g. GLAD alerts in the Amazon) or dropped automatically (e.g. GLAD alerts on forestry areas).
|Biome||Monitoring systems||Native vegetation class(es)|
|Amazon||DETER-B-Amazônia e SAD||Forest Formation|
|Cerrado||DETER-Cerrado||Forest Formation; Savanna Formation; Grassland Formation|
|Caatinga||GLAD-Alerts||Forest Formation; Savanna Formation; Grassland Formation|
|Pampa||GLAD-Alerts||Forest Formation; Wetland; Grassland Formation|
|Pantanal||GLAD-Alerts||Forest Formation; Savanna Formation; Grassland Formation; Other Non-Forest Formation|
|Atlantic Forest||GLAD-Alerts||Forest Formation; Mangrove; Savanna Formation; Grassland Formation|
- Pre-activation: Stage operated by experts to identify which of the alerts should be removed because they are false positives of deforestation, being performed by visual inspection of public images (e.g. Sentinel) and/or free geo-services (e.g. Planet Monthly Mosaics). The main objective of this step is to remove the alerts that are not real, thus avoiding the unnecessary activation of Planet images for regions with false alerts of deforestation.
- Planet Activation: The stage responsible for activating/acquiring Planet images for regions with valid alerts, according to the experts' review, for further refinement of the spatial delimitation of these deforestation alerts. The activation of the images occurs via WEB services of an API developed by Santiago and Cintra Consultoria-SCCON (Planet's representative in Brazil), which cuts the requested images, according to the regular grid of 5x5km that covers the entire Brazilian territory, and stores this result in the Google Cloud Storage platform. The Planet images are stored with all spectral bands (e.g. blue, green, red and near infrared), unusable data mask (e.g. Unusable Data Mask-UDM) and respective metadata.
- Planet Ingestion: At this stage, the activated Planet images are ingested on the Google Earth Engine (GEE) platform via the Python API. The storage of images in Google Cloud Storage is critical since Google Earth Engine has a native and efficient integration with this storage platform. The GHG will refine the alerts and identify the images before and after the deforestation event.
- Refinement of Alert: Stage responsible for refining the alerts of deforestation with reference to the activated Planet images. The refinement process begins when all the pre-selected alerts and their respective Planet images are available in the Toolkit-MapBiomas, a tool that facilitates the operation of experts in Google Earth Engine, as described below: Selection of the alert that will be refined; Choice of the two best Planet images, before and after the deforestation; Delimitation of the region of interest based on the observed deforestation; Delimitation of deforestation and non-deforestation samples within the region of interest; Redesign of the alert polygonal containing the effective area of deforestation, through supervised classification using the selected samples and the Random Forest algorithm; Fine adjustment of the polygonal layout resulting from the classification.
- Export: At this stage the refined alerts, the samples used for classification of the final polygonal of the alert, and their respective Planet images, pre- and post-deforestation, are exported to the MapBiomas-Alerta system, hosted on the Google Cloud Platform.
- Spatial Analysis: with operations within the database, spatial crossings of alerts with various thematic maps (e.g. CAR, Conservation Units, Indigenous Lands, Settlements) are performed and are available for the Dashboard, APIs for remote access and report generation.
- Publication of Reports: For each CAR that contains an alert is produced a Report containing the (i) ID of the CAR, (ii) source of the alert, (iii) images of before and after deforestation, (iv) location of the property and location of the alert within the property, (v) location of the alert / ownership in the UF, (vi) data overlap with IT, UC, settlements, legal reserve, APP and springs, (vii) existence of an embargo, management plan or authorization to suppress the property, (viii) historical coverage of the area in previous years (based on the MapBiomas Collection) and (ix) descriptive memorial of the alert area on the property. In the case of alerts that do not cross CAR properties, a simplified report is generated with items (ii), (iii), (v), (vi), (vii), (viii) and (ix).
- Publication on the Dashboard: all alerts >0.2 ha are published on the online platform (http://alerta.mapbiomas.org) organized in such a way as to enable the visualization of the alerts on the map and the statistics of the number of alerts and deforested area by biome, state, and municipality for the selected period. Each alert can be viewed with images from before and after the deforestation with date and area affected and a link to the related reports. In a restricted access area, dashboards are mounted for consultation of reports and customized reports for different public agencies or public interest (ex IBAMA).
- Publication in Service APIs: APIs are built for consultation and access to alerts and customized reports for different public and public interest agencies.